The analysis of time series data is widely used to characterize the behavior of a system. The following four general categories of approaches are commonly applied to achieve characterization of such a system and these provide a general background for the present invention. The approaches are illustrative both in their conceptualization, application, and limitations.
The first such approach represents a form of mathematical reductionism of the complexity through the application of a cascade of rules based on an anticipated relationship between the time series output and a given set of system mechanisms. In this approach the operative mechanisms, data set characteristics, and intruding artifact are a priori defined to the best extent possible. Then a set of rules is applied to characterize systems and analyze the data set based on predicted relationships between the data set and the systems being characterized. Such systems often include cascading branches of decision-based algorithms, the complexity of which increase greatly in the presence of multiple interactive mechanisms. The reductionism approach is severely limited by the uncertainty and complexity, which rapidly emerges when a cascade of rules is applied to a highly interactive data set, when the signal to noise ratio is low, and/or when multiple data sets generated by complex and dynamically interactive systems are evaluated. These methods become inordinately more cumbersome as the complexity and number of time series increases. In addition the subtlety of the interactive and dynamic relationships along and between datasets and the variations associated with the technique or tools of data collection often makes the cascading rules very difficult to a priori define.
The failure of simplification the analysis through mathematical reductionism to adequately characterize the complex systems generating such data sets, led to the perception that this failure resulted from specific limitations of a particular data format (usually the time domain format). In other words the time series was perceived to contain sufficient information to characterize the system but, it was thought, that the recognition of this information required reformatting into a different mathematical representation, which emphasized other hidden components which were specific for certain important system characteristics. This approach is exemplified by frequency processing methods, which reformat the time series into frequency components, such as its sine components or wavelets, with the hope that patterns of specific frequency relationships within the system will emerge to be recognized. While often uncovering considerable useful information, this approach is remains quite limited when applied to highly complex and interactive systems, because many complex relationships are poorly characterized by their frequency components, and it is often difficult to relate an output derived from frequency-based primitives to specific mechanisms operative within the system. In other words, the advantages associated with mathematically defined linkages between system mechanisms and the rules based analysis provided by reductionism is reduced by the data reformatting process for the purpose of frequency based signal processing as, for example, is provided by Fourier or wavelet transforms.
A third approach seeks to identify the patterns or relationships by repetitively reprocessing the time series with a set of general comparative rules or by statistical processing. As with the data reformatting approach the utility of this method in isolation (as embodied in neural network based analysis), is severely limited by dissociation of the output from the complex and interactive operative mechanisms, which define the output. With such processing, the relevant scope and characterization of the relationships of the output to the actual behavior of the dynamic interactions of the system is often quite limited. This limits the applicability of such processing in environments wherein the characterization of behavior of the system as a function by the output may be as important as the actual output values themselves.
A fourth approach has been to apply chaotic processing to the time series. Again, like that of conventional signal processing this alternative method is applied the expectation that some predictive pattern will emerge to be recognized. This technique shares several of the limitations noted for both frequency and statistical based data reformatting. In addition as will be discussed the application of this type of processing to physiologic signals is limited by redundant and interactive higher control which greatly limits the progression of the system to a state of uncontrolled chaotic behavior. Such systems operate in environments of substantial interactive control until the development of a severe disease state a point at which the diagnostic information provided by processing often has less adjective utility relevant timely intervention.
The human physiologic system derives a large array of time series outputs, which have substantial relevance when monitored over a finite time interval. The human can be considered the prototypic complex interactive system. These interactions and the mechanisms defining them have been the subject of intense research for over one hundred years and most of this work has been performed the time domain. For this reason any approach toward the characterization of such a system needs to consider the value of engaging the body of knot ledge which relates to these mechanisms. This has been one of the reasons that the reductionism has predominated in the analysis of physiologic signals. U.S. Pat. No. 5,765,563 to Vander Schaff, U.S. Pat. No. 5,803,066 to Rapoport, and U.S. Pat. No. 6,138,675 to Berthon-Jones show such simple cascade decision systems for processing physiologic signals. U.S. Pat. No. 5,751,911 to Goldman shows a real-time waveform analysis system, which utilizes neural networks to perform various stages of the analysis. U.S. Pat. No. 6,144,877 to Depetrillo shows a processor based method for determining statistical information for time series data and for detecting a biological condition of a biological system from the statistical information. U.S. Pat. Nos. 5,782,240 and 5,730,144 to Katz shows a system, which apply chaos analysers, which generate a time series, vector representation of each monitored function and apply chaotic processing to identify certain events. All of these systems are deficient in that they are not able to adequately organize, order and analyze the true state of dynamic interaction operative in the generation of these signals.
Critical illness is one example of a dynamic timed process, which is poorly characterized by the above noted conventional methods. When human physiologic stability is under threat, it is maintained by a complex array of interactive physiologic systems, which control the critical time dependent process of oxygen delivery to the organism. Each system (e.g. respiratory, cardiac or vascular) has multiple biochemical and/or mechanical controls, which operate together in a predictable manner to optimize oxygen delivery under conditions of threat. For example, an increased oxygen requirement during infection causes the patient to increase oxygen delivery by lowering lung carbon dioxide through hyperventilation and the fall in carbon dioxide then causes the hemoglobin molecule to increase its affinity for oxygen thereby further enhancing oxygen delivery. In addition to the basic control of a single system, other systems interact with the originally affected system to producing a predictable pattern of response. For example, in the presence of infection, the cardiac system interacts with the respiratory system such that both the stroke volume and heart rate increase. In addition, the vascular system may respond with a reduction in arterial tone and an increase in venous tone, thereby both reducing impedance to the flow of oxygen to the tissues and shifting more blood into the arterial compartment.
Each system generally also has a plurality of predicable compensation responses to adjust for pathologic alteration or injury to the system and these responses interact between systems. For example the development of infectious injury to the lung will result in an increase in volume of ventilated gas to compensate for the loss of functional surface area. This increase in ventilation can then induce a synergistic increase in both stroke volume and heart rate.
Finally a pathologic process altering one system will generally also induce an alteration in one or more other systems and these processes are all time dependent. Sub acute or acute life threatening conditions such as sepsis, pulmonary embolism, or hemorrhage generally affect the systems in cascades or predictable sequences which may have a time course range from as little as 20 seconds or more than 72 hours. For example, the brief development of airway collapse induces a fall in oxygen saturation, which then causes a compensatory hyperventilation response which causes a rise in heart rate over as little as 20-30 seconds. An infection, on the other hand, has a more prolonged time course inducing a rise in respiration rate a rise, in heart rate, and then a progressive fall in oxygen saturation and finally a fall in respiration rate and a finally a terminal fall in heart rate often over a course of 48-72 hours.
It can be seen therefore that each disease process engaging the organism causes the induction of a complex and interactive time series of pathophysiologic perturbation and compensation. At the onset of the disease (such as early in the course of infection) the degree of physiologic change may be very slight and limited to one or two variables. As a disease progresses both the magnitude of perturbation and the number of system involved increases. In addition to inducing a predictable range of perturbation, a particular disease process generally produces a specific range of progression and pattern of evolution as a function of injury, compensation, and system interaction. Furthermore, this multi-system complexity which can be induced by initial pathologic involvement of a single system is greatly magnified when a plurality of pathologic processes is present.
Despite the fact that these conditions represent some of the most important adversities affecting human beings these pathologic processes are poorly characterized by even the most sophisticated of conventional monitors, which greatly oversimplify the processing and outputs. Perhaps this is due to the fact that this interactive complexity overwhelmed the developers of substantially all of the conventional physiologic signal-processing methods in the same a that it overwhelms the physicians and nurses at the bedside everyday. Hospital critical care patient monitors have generally been applied as warning devices upon threshold breach of specific critical parameters with the focus on the balance between timely warning of a potentially life threatening threshold breach and the mitigation of false alarms. However, during the pivotal time, early in the process of the evolution of critical illness, the compensatory responses limit the change in primary critical variables so that the user, monitoring these parameters in isolation, is often given a false sense of security. For this reason, it cannot be enough to recognize and warn of the occurrence of a respiratory arrest, or hypotension, or hypoxia, or of a particular type of cardiac arrhythmia. To truly engage and characterize the processes present, a patient monitor must have capability to properly analyze, organize, and output in a quickly and easily understood format the true interactive state of critical illness. As discussed below, it is one of the purposes of the present invention to provide such a monitor